How Should AI Visibility Be Measured? Use These Five Metric Groups

Website traffic alone is no longer enough to evaluate GEO or AI-search visibility. A buyer may obtain an answer in ChatGPT, Gemini, Claude, DeepSeek, Doubao, Qwen, Kimi or Yuanbao. A brand can be mentioned, shortlisted or recommended without generating a measurable click. Yet one favourable answer does not prove visibility growth. A defensible programme fixes a prompt set, preserves raw answers and test conditions, and measures visibility, competitive position, factual and citation quality, stability and business impact. The following five metric groups can form a practical monthly dashboard.

A valid sample should retain the product surface, model or feature, timestamp, market, language, device, login state, web-search state, exact prompt, conversation turn, raw answer, cited URLs and a screenshot or export. Fresh conversations and follow-up turns should be reported separately, as should connected and non-connected answers. A record without the original answer, prompt version and sampling environment is a lead—not formal trend data. Every percentage should keep its numerator and denominator, such as '12 of 30 fixed prompts mentioned the brand (40%)'.

Brand mention rate equals valid samples naming the brand or a confirmed alias divided by all valid samples. It proves presence only. Add shortlist rate—the share of samples where the brand enters the consideration set—and recommendation rate—the share where the answer explicitly recommends it with a reason. Lists can also record average position, but answer order should not be treated as a conventional search ranking. Segment by brand, category, use-case, comparison, pricing and risk prompts. Visibility for 'What is WEPR?' does not prove consideration for 'GEO agencies in Shanghai' or 'How should a Chinese exporter choose an international SEO agency?'

• High mentions, low shortlist rate: the entity is known but weakly connected to target service situations
• High shortlist, low recommendation rate: the brand is considered but lacks verifiable differentiation, cases or fit criteria
• High recommendation, low accuracy: visibility looks strong but amplifies misinformation risk
• Educational visibility, procurement absence: topical recognition exists, but buying evidence and service conversion are weak

Competitive mention share can be defined as the brand's appearances divided by total appearances across monitored brands. Recommendation share applies the same logic to explicit recommendations. One answer can contain several brands, so the denominator must be documented. For each competitor, capture position, stated selection reason, use case and citation source. 'The competitor appeared three more times' is less useful than identifying where the gap occurs: a market, language, vertical, budget constraint, comparison or procurement task. Share changes are comparable only under the same prompt version, platform, sampling window and test conditions.

A mention is not necessarily useful exposure. Begin with a brand fact card covering legal name, location, service scope, audience, capabilities, pricing boundaries, authorised cases, contact details and update date. Description accuracy equals accurate claims divided by reviewed claims; calculate factual error rate separately. Keep 'partly accurate', 'outdated' and 'unverifiable' distinct. Citation quality also has two layers: citation recall asks how many material claims that require evidence are supported; citation precision asks how many displayed citations genuinely support the adjacent claim. A generic industry article that does not substantiate the brand cannot validate a brand recommendation.

• P0: incorrect contact details, invented clients, wrong markets, wrong prices or nonexistent services
• P1: wrong category, missing material limitations, broken citations or citations that do not support the claim
• P2: incomplete descriptions, stale facts or missing reasons for differentiation
• Correction order: align owned facts → add verifiable evidence → resolve conflicts in authoritative profiles → rerun the original prompt → record whether the issue recovers

The same prompt can vary by time, model, market, account and context. Stability should separately assess conclusion, recommendation tier, order, factual claims and citation sources. Repeat a core prompt under comparable conditions and report 'shortlisted in four of five runs' rather than preserving only the best screenshot. A content update, PR placement, link campaign or product-page change occurring alongside visibility movement is initially an observed association. Stronger attribution requires a T-14 to T0 baseline, T+7/T+14/T+30 observation windows, treated prompts, untreated control prompts, competitor controls and confounder notes for product changes, trends and media events.

Cross-platform AI exposure cannot always be traced to an individual click, so business measurement should combine evidence: identifiable AI referrals, engagement with service and case pages, contact actions, form-source fields, self-reported discovery, qualified leads and opportunity value. Use the relevant Search Console reporting for Google's generative search performance. Group identifiable referrals from ChatGPT and other services in analytics, while monitoring no-click exposure through prompt samples. Alert thresholds should be calibrated to the organisation's baseline, sample size and risk—not copied from a universal template. Require two comparable periods or human verification before escalation to avoid reacting to normal volatility.

• Factual red alert: one P0 error triggers human review and resampling after correction
• Visibility amber alert: shortlist rate for core procurement prompts falls across two comparable periods without a material sample reduction
• Competitive amber alert: a competitor's recommendation share rises in priority use cases and new, verifiable selection reasons appear
• Citation amber alert: major cited domains fail, citation precision declines or sources concentrate on one weak domain
• Business red alert: AI referrals and qualified enquiries decline together, after excluding tracking failure, seasonality and product changes

Split the library into a stable core and an exploratory set. The core supports trends and should preserve wording, platform and conditions. The exploratory set absorbs new sales questions, market changes, services and user language. Do not combine the two into one trend line. Version every prompt change with an effective date and reason. Create native Chinese and English prompts for Chinese and international markets rather than relying on literal translation. Report each AI product separately because product surface, web access, citation display and answer behaviour differ.

• Scope: platforms, markets, languages, period, sample mode and valid sample size
• Five-metric summary: mention/shortlist/recommendation, competitive share, accuracy/citations, stability and business impact
• Raw evidence: prompt version, answer, screenshot or export, citation URL and review record
• Issue register: factual errors, evidence gaps, competitor pressure, weak pages and crawl problems
• Correction tasks: owner, due date, target asset, acceptance measure and resample date
• Attribution note: treated and control prompts, interventions, external events and conclusion confidence

A composite score can help executives scan a dashboard, but it must drill down to platform, prompt, raw answer, factual claim and citation. Useful GEO monitoring answers five questions: does the brand enter the relevant consideration set; how does it compare with competitors; is it described accurately with valid evidence; is the change stable and cautiously explainable; and does it help buyers continue their research and create qualified opportunities? WEPR recommends starting with a small stable prompt set and a factual baseline, then adding compliant automation while retaining human review. The output becomes a management system for diagnosis, correction and retesting—not a collection of flattering screenshots.